trace: make window materialization atomic

tests/test_core_flow.py

@@ -853,6 +853,64 @@ class CoreFlowTests(unittest.TestCase):
             self.assertEqual(rows[0]["output_length"], 7)
             self.assertIsInstance(rows[0]["sampling_u"], float)
 
+    def test_prepare_trace_windows_preserves_existing_files_on_failure(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            tmp_path = Path(tmp)
+            legacy_source = tmp_path / "legacy"
+            thinking_source = tmp_path / "thinking"
+            output_root = tmp_path / "trace_windows"
+            traces_dir = output_root / "traces"
+            legacy_source.mkdir()
+            thinking_source.mkdir()
+            traces_dir.mkdir(parents=True)
+
+            for filename in [
+                "qwen_chat_blksz_64_031109-031111",
+                "qwen_chat_blksz_64_031121-031123",
+            ]:
+                for suffix in [".jsonl", "_prompt.jsonl"]:
+                    path = legacy_source / f"{filename}{suffix}"
+                    path.write_text(
+                        json.dumps(
+                            {
+                                "chat_id": "c1",
+                                "turn": 1,
+                                "timestamp": 3605.0,
+                                "input_length": 20,
+                                "output_length": 7,
+                                "prompt": "prompt",
+                            }
+                        )
+                        + "\n",
+                        encoding="utf-8",
+                    )
+
+            sentinel = traces_dir / "chat_w20260311_1000.jsonl"
+            sentinel.write_text("sentinel\n", encoding="utf-8")
+
+            proc = subprocess.run(
+                [
+                    "python3",
+                    "scripts/prepare_trace_windows.py",
+                    "--legacy-source",
+                    str(legacy_source),
+                    "--thinking-source",
+                    str(thinking_source),
+                    "--output-root",
+                    str(output_root),
+                    "--workloads",
+                    "chat",
+                    "--overwrite",
+                ],
+                cwd="/home/gahow/phd/aituner",
+                capture_output=True,
+                text=True,
+            )
+
+            self.assertNotEqual(proc.returncode, 0)
+            self.assertEqual(sentinel.read_text(encoding="utf-8"), "sentinel\n")
+            self.assertEqual(sorted(path.name for path in traces_dir.glob("*.tmp.*")), [])
+
     def test_binary_search_max_feasible(self) -> None:
         result = binary_search_max_feasible(
             low=0.0,